jonigl/mcp-client-for-ollama

uv tool install --upgrade ollmcp
pip install --upgrade ollmcp
uvx ollmcp (run without installing)

Usage

• •Register a server: ollmcp mcp add playwright -- npx @playwright/mcp@latest (scopes: local/project/user)
• •Run the client: ollmcp (use /help for interactive commands)
• •Connect remote servers: ollmcp --mcp-server-url http://localhost:8000/sse
• •Pick provider/model: ollmcp --provider openai --api-key $OPENAI_API_KEY --model gpt-5.5

Key features

• ✓Multi-server support over stdio, SSE, and Streamable HTTP transports
• ✓Agent Mode: iterative tool execution with configurable loop limit (default 7) and choose-your-own-adventure options at the limit
• ✓Human-in-the-Loop approval before tool execution (enabled by default; session/abort/disable options)
• ✓Ollama Cloud models plus OpenAI-compatible providers (OpenAI, OpenRouter, DeepSeek, etc.), settings remembered per provider
• ✓MCP Prompts (text-only) and MCP Resources (text + images) with autocomplete; thinking mode and reasoning-effort levels
• ✓Advanced model config (15+ params), vision tool support, history export/import, server hot-reload for development

Best for

Harness engineering and local-first agent development — steering local Ollama models with MCP tools (and HIL safety) in a controlled terminal.

Caveats

• ⚠Requires Python 3.11+, Ollama running locally, and (recommended) the uv package manager
• ⚠Non-Ollama providers are experimental; capability detection (tools/vision/thinking) is assumed for them and may be wrong
• ⚠MCP Prompts support text content only — images, audio, and embedded resources are skipped
• ⚠Breaking change: interactive commands now require a leading /
• ⚠Only keys passed with --api-key are stored in plain text in ~/.config/ollmcp/config.json